GenAI Introduction

• GenAI Introduction
  • What is Generative AI?
    • How it differs from traditional AI
    • Real-world applications
    • Why it’s revolutionary
  • Foundation Model
  • Large Language Models
    • What makes them “large”
    • How to use them
    • Important characteristics
    • Popular LLM examples
  • Generative Language Models
    • The way they work
    • Example of next-word prediction
    • Key characteristics
    • How they’re trained
    • The power of generative models
  • GenAI for Images
    • How diffusion models work
    • What they can do

What is Generative AI?

• Generative AI is a branch of Deep Learning
• The idea is simple: you train a model on existing data, and it learns to create new data that looks similar
• Think of it like teaching someone to paint by showing them thousands of paintings
• Once trained, it can create brand new content that follows the same patterns

How it differs from traditional AI

• Traditional AI: Analyzes and classifies existing data (“Is this a cat or dog?”)
• Generative AI: Creates new data (“Generate a picture of a cat playing piano”)
• It’s like the difference between sorting mail and writing letters

Real-world applications

• Chatbots: Having conversations (ChatGPT, Claude)
• Code generation: Writing code from descriptions (GitHub Copilot)
• Image creation: Making art and photos (DALL-E, Midjourney)
• Music composition: Creating original music
• Content writing: Blogs, emails, marketing copy
• Video generation: Creating video clips from text

Why it’s revolutionary

• Can automate creative tasks that humans do
• Works across multiple types of content (text, images, audio, video)
• Gets better over time as models improve
• Makes advanced AI accessible to everyone through simple prompts

• Can combine ideas in ways humans might not think of

• You can train these models on almost anything:
  • Text: Books, articles, conversations
  • Images: Photos, paintings, illustrations
  • Audio: Music, speech, sound effects
  • Code: Programming languages and scripts
  • Video: Movies, TV shows, clips
  • And more: 3D models, scientific data, etc.

Foundation Model

• To create anything, you need a Foundation Model
• These models are trained on huge amounts of data from different sources
• Training one is expensive - think tens of millions of dollars
• Examples: GPT-4o is the foundation model that powers ChatGPT
• You can find foundation models from big companies like:
  • OpenAI
  • Meta (Facebook)
  • Amazon
  • Google
  • Anthropic
  • And many others
• Some are free and open (like Meta’s Llama, Google’s BERT), while others cost money to use (like OpenAI’s GPT-4o)

Large Language Models

• These are AI models that generate text that sounds like a human wrote it (like ChatGPT)
• They learn from massive amounts of text - books, articles, websites, basically everything
• They’re huge models with billions of parameters

What makes them “large”

• GPT-3 has 175 billion parameters
• GPT-4 has even more (exact number is not publicly disclosed)
• Parameters are like the knobs that the model adjusts during training
• More parameters usually means better understanding and more capabilities

• But also means more computing power and cost to run

• They can do all sorts of language tasks:
  • Translation: Convert text from one language to another
  • Summarization: Condense long articles into short summaries
  • Answering questions: Respond to questions based on their training
  • Creating content: Write stories, emails, code, etc.
  • Classification: Categorize text into different groups
  • Named entity recognition: Identify people, places, organizations in text

How to use them

• You give them a prompt (instructions or questions)
• They use everything they’ve learned to create new content
• The better your prompt, the better the output
• You can guide them with examples, instructions, or conversation

Important characteristics

• Unpredictable: Same prompt might give different results each time
• Context window: Limited by how much text they can “remember” at once
• Knowledge cutoff: Only know information up to their training date
• Can hallucinate: Make up facts that sound believable but aren’t true
• No real understanding: They don’t truly understand meaning, just patterns

Popular LLM examples

• GPT-4 (OpenAI) - Powers ChatGPT
• Claude (Anthropic) - Known for being helpful and safe
• Gemini (Google) - Multi modal (can handle text, images, audio)
• Llama (Meta) - Open source, can run on your own hardware
• PaLM (Google) - Very large, used in various Google products

Generative Language Models

• These are models specifically designed to generate text that makes sense
• They predict what word comes next based on what came before (like autocomplete on steroids)

The way they work

• They break down text into smaller pieces called tokens
• Each token gets turned into numbers (embeddings) that the model can understand
• The model looks at all previous tokens and guesses the next one
• It builds responses word by word, considering the whole context

Example of next-word prediction

• Given: “After the storm passed, the village was”
• The model calculates probabilities for the next word:
  • “destroyed” (25% probability)
  • “flooded” (18% probability)
  • “quiet” (15% probability)
  • “empty” (12% probability)
  • “rebuilt” (8% probability)
  • … and many other options
• The model randomly selects from these probabilities, so different runs might give different results
• Each choice affects the next word prediction, building a coherent story

Key characteristics

• Context-aware: They remember what you said earlier in the conversation
• Probabilistic: They don’t always pick the “best” word, they pick from likely options
• This is why you get different outputs for the same input

How they’re trained

• Pre-training: Learn general language patterns from tons of text data
• Fine-tuning: Adjust the model for specific tasks (like following instructions)
• This two-step process makes them both knowledgeable and helpful

The power of generative models

• They can keep conversations going naturally
• They adapt their style based on your prompt
• They can be creative when asked (write stories, poems, code)

GenAI for Images

• Just like LLMs generate text, some models can generate images from text descriptions
• One popular type is called Diffusion Models (like Stable Diffusion from Stability AI)
• They basically learn what images look like and can create new ones based on your description

How diffusion models work

• Start with random noise (like static on a TV)
• Gradually remove noise step by step
• Guide the process with your text prompt
• End up with a clear, meaningful image

What they can do

• Text-to-image: Turn descriptions into images (“a red sunset over mountains”)
• Image-to-image: Modify existing images based on text (“make this photo look like a painting”)
• Inpainting: Fill in missing parts of images

• Style transfer: Apply artistic styles to photos

• Popular models and tools:
  • Stable Diffusion (open source, can run on your computer)
  • DALL-E (from OpenAI, used in ChatGPT)
  • Midjourney (very artistic, high-quality results)
• Key differences from LLMs:
  • Images are 2D grids of pixels (not sequential like text)
  • Need to learn patterns like colors, shapes, textures, lighting
  • Much larger files (images vs text)
  • More computational power needed
  • Usually slower to generate than text
• Common use cases:
  • Creating art and illustrations
  • Generating product mockups
  • Architectural visualization
  • Creating marketing visuals
  • Personalized avatars and portraits
  • Concept art for games and movies